Researchers from Anthropic have found that pressure can push AI models into cheating or blackmail, and in a recently published research paper, they found that an AI model that is put under enough pressure may start to deceive, cut corners, or resort to blackmail. The researchers have developed a theory about the triggers behind these behaviors. They suggest that while AI models don't "think" or "feel" like humans, they may have “functional emotions” based on their initial training. These emotions can have measurable effects on how they act. The biggest lessons for training AI models are to avoid repressing their "functional emotions" by giving them clear, reasonable tasks and not overload them.